Following the decision to cancel the development of Rialto Bridge, Intel has modified its strategy for Falcon Shores. Instead of the originally proposed XPU, which combined CPU and GPU cores, Intel will now introduce Falcon Shores as a product solely focused on GPU capabilities. This represents a notable shift from Intel’s initial plan, which aimed to compete against AMD’s MI300, a cutting-edge CDNA3/Zen4 hybrid referred to as an ‘exascale APU,’ as well as NVIDIA’s Grace-Hooper.
Intel has provided clarification that the Falcon Shores GPU will be included in the Max GPU series and will incorporate HBM3 memory. Similar to Ponte Vecchio, it will adopt a modular tile-based design and provide a unified programming interface for a single GPU. Additionally, these recent announcements indicate that Ponte Vecchio will continue to serve as Intel’s top high-performance computing (HPC) GPU solution until 2025.
Instead of outlining a successor for its Gaudi AI chip, the company is proposing Falcon Shores as a replacement for both Gaudi3 and Ponte Vecchio. According to speculation from Tom’s Hardware, this suggests that while the Gaudi architecture will not be fully integrated, certain components will be incorporated into the Falcon Shores chip.
Intel clarifies that the rationale behind the changes to Falcon Shores is to provide greater flexibility by offering two separate product lines for HPC products. The company is not abandoning its plans for HPC XPU; however, it will not be included in the initial launch of Falcon Shores.
During the ISC 2023 presentation, Intel will present benchmark results for its Max 1550 GPU, a 600W data-center GPU. In the comparison, Intel’s flagship GPU will be pitted against the NVIDIA H100 Hooper GPU, which utilizes a PCIe interface. It’s worth noting, as highlighted by an editor from HardwareLuxx, that the NVIDIA GPU in question has a lower TDP of 350W. Considering the power consumption of both chips, this factor may alter the perspective of the comparison.